I have recently come across comments from people working on machine and statistical learning referring to supervised learning as the way in which we learned concepts at an early age. These people often cite examples such as “when you were a toddler, your parents showed you pictures of houses and that’s how you learned what a house was”. They cite examples of this sort as if suddenly by magic this particular toddler started seeing houses where before they saw nothing. This is total nonsense. The only knowledge that the toddler gained after being shown examples of houses and non-houses was the ability to call a house “a house”, that is, they only learned the name by which an already learned concept is called within their mother language.
A toddler knows the concept of a house, and thus of a non-house, before they can name it. They know that the place where they spend most of their day, where they sleep and are fed, is different from the place where their parents take them every morning where they are in company of other toddlers and in the care of people that are not their parents. They know that the place where they spend most of their day and where they feel safe is similar to that place where their parents take them once in a while where they meet these old human beings that treat them with love and care. A toddler learns the concept of a house, and therefore of home, in an unsupervised fashion, by being repetitively exposed to the experience of being in houses and non-houses. The mechanisms of such unsupervised learning is something that we don’t know yet.
However, I believe that supervised learning becomes important later in life where we acquire a language and a solid conceptual mental life in order to learn novel concepts that do not rely on direct experience. But even there the question of how we perform supervision (i.e. the supervised part in supervised learning) differs from what is used in artificial intelligence. Or which of these two approaches does the reader think is a better way to learn what a Kwijibo is? Approach #1: I show you one thousand images of a Kwijibo in different scenarios where each image is labeled as “Kwijibo”, and mixed with other one thousand images of any other animal where each image is labeled as “not-Kwijibo”, and then expect that you learn the commonalities and differences among those two thousand exposures to Kwijibos and non-Kwijibos, or… Approach #2: I rely on your knowledge of language (and thus, of concepts) and simply tell you that a Kwijibo is a fat, balding north-american ape. Which of those is a more efficient way of learning? Approach #2 becomes our standard idea of supervised learning when afterwards you show me a picture of an orangutan and ask me whether I meant that when I told you what a Kwijibo was; for which I would reply negatively, and you would then readjust your concept.
Having said that, I strongly believe that in order to reach artificial general intelligence we should focus on unsupervised learning algorithms along with hybrid methods that combine the power of artificial neural networks with symbolic (that is, related to concepts) approaches.
The Thoughtist 